Electronic element for reservoir calculation

Abstract

To provide an electronic element for reservoir calculation, which has high nonlinearity in a reservoir portion, can suppress calculation resources, is compact, and is suitable for integration.SOLUTION: An electronic element (reservoir element, electric double layer transistor, EDLT) 101 includes a semiconductor layer (hydrogen-terminated diamond) 11, an ionic conductor (LSZO) 16, a gate electrode 14, a source electrode 12, and a drain electrode 13. The source and drain electrodes are in electrical contact with the semiconductor layer. The gate electrode is arranged via an ionic conductor in a channel layer 15 formed between an electrical contact surface of the source electrode and an electrical contact surface of the drain electrode on the semiconductor layer.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an electronic device for reservoir computing.

Background Art

[0002] In recent years, neural networks and artificial intelligence that mimic the brain neural network have been actively studied. For example, artificial intelligence technologies using multi-layer neural networks including deep learning are known. Although this technology generally has high performance, it has problems such as requiring a very large amount of computing resources, consuming a large amount of power, and having a large equipment size. Therefore, there is a high barrier to its application to small portable terminals with limited power and volume (size). [[ID=...]]

[0003] As another approach to neural networks, there is reservoir computing (reservoir computing), which is disclosed in, for example, Patent Documents 1 to 3. Reservoir computing uses the non-linearity of physical phenomena for computing and can, in principle, perform computations with few resources. However, conventionally, the non-linearity (expressive power) of the reservoir part has been low, and sufficient reduction of computing resources has not been possible. For this reason, there have been problems such as a large size of the reservoir part and difficulty in high integration.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Patent Document 3

Summary of the Invention

Problems to be Solved by the Invention

[0005] An object of the present invention is to provide an electronic device for reservoir computing that has a high non-linearity in the reservoir section, can suppress calculation resources, and is compact and suitable for integration.

Means for Solving the Problems

[0006] The configuration of the present invention for solving the problems is shown below. (Configuration 1) An electronic device for reservoir computing having a semiconductor layer, an ion conductor, a gate electrode, a source electrode, and a drain electrode, where the source electrode and the drain electrode are in electrical contact with the semiconductor layer, and the gate electrode is disposed via the ion conductor in a channel layer formed between an electrical contact surface of the source electrode and an electrical contact surface of the drain electrode on the semiconductor layer. (Configuration 2) The semiconductor layer of the electronic device for reservoir computing according to Configuration 1 is made of one selected from the group consisting of a hydrogen-terminated diamond semiconductor, silicon, GaAs, Ga2O5, SiC, GaP, InP, ZnSe, CdS, GaN, SiGe, CuInSe2, graphene, monolayer MoS2, and ZnO. (Configuration 3) The semiconductor layer of the electronic device for reservoir computing according to Configuration 2 is made of a hydrogen-terminated diamond semiconductor. (Configuration 4) The conductor ions of the ion conductor are Li + ions, H + ions, Na + ions, Ag + ions, Cu + ions, Cl - ions, Br - ions, I - ions, SO4 - ions, FeCl4 - ions, OH - ions, BF4 - ions, PF6 - ions, CF3SO - ions, (FSO2)2N- Ion, (CF3SO2)2N - Ion, C6H 11 N2 + Ion, C 10 H 20 F6N2O5S2 + An electronic element for reservoir calculations, comprising one or more elements selected from the group consisting of ions, as described in any one of the configurations 1 to 3. (Composition 5) The conductor ions of the aforementioned ion conductor are Li + An ion-based electronic element for reservoir calculations, as described in configuration 4. (Composition 6) The ion conductor is in solid form, and the reservoir calculation electronic element is as described in any one of the configurations 1 to 5. (Composition 7) The ion conductor is gel-like, and the reservoir calculation electronic element is as described in any one of items 1 to 5. (Composition 8) The ion conductor is in a liquid state, and the reservoir calculation electronic element is as described in any one of items 1 to 5. (Composition 9) The ionic conductor is an electronic element for reservoir calculations according to any one of the configurations 1 to 6, consisting of Li-SiO2-ZrO2(LZSO). (Composition 10) The gate electrode, source electrode, and drain electrode are made of one or more metals selected from the group consisting of aluminum (Al), tungsten (W), gold (Au), titanium (Ti), platinum (Pt), palladium (Pd), copper (Cu), silver (Ag), magnesium (Mg), zinc (Zn), iron (Fe), cobalt (Co), nickel (Ni), and molybdenum (Mo), an alloy containing one or more metals selected from the group, doped polysilicon, carbon, graphene, and graphite, respectively, for a reservoir calculation electronic element according to any one of configurations 1 to 9. (Composition 11) The reservoir calculation electronic element according to any one of configurations 1 to 10, wherein the channel layer has multiple channels with different channel widths arranged therein. [Effects of the Invention]

[0007] According to the present invention, an electronic element for reservoir calculations is provided that has high nonlinearity in the reservoir section, can reduce computational resources, and is compact and suitable for integration. [Brief explanation of the drawing]

[0008] [Figure 1] This is a cross-sectional view showing the structure of the electronic element of the present invention. [Figure 2] This is an explanatory diagram illustrating the operation of the electronic element of the present invention. [Figure 3] This is a characteristic curve showing the drain current response when time-series data ("1011") is input. [Figure 4] This is a characteristic diagram showing the reservoir state when 16 different pixel information values ​​are input. [Figure 5] This is a characteristic diagram showing the relationship between the number of training images and the image recognition accuracy. [Figure 6] This characteristic diagram shows the relationship between the input waveform u(κ) and y(κ) expressed by equation (1) for discrete time κ. [Figure 7] This is a characteristic diagram showing the agreement between the predicted waveform and the target waveform. [Figure 8] This is an explanatory diagram illustrating the structure and operation of the electronic element of the second embodiment. [Figure 9] This is a characteristic diagram showing the agreement between the predicted waveform and the target waveform in Example 2. [Figure 10] This is a characteristic diagram showing the agreement between the predicted waveform and the target waveform in Example 2. [Figure 11] This is a diagram showing the configuration of an information processing device. [Modes for carrying out the invention]

[0009] The embodiments for carrying out the present invention will be described below with reference to the drawings. Note that the expressions A to B appearing in the text indicate A or greater and B or less.

[0010] (Embodiment 1) Embodiment 1 describes an electronic element for reservoir calculations using an electric double-layer transistor with an ion conductor and a semiconductor.

[0011] <Structure of electronic elements> As shown in Figure 1(a) illustrating its main configuration, the reservoir calculation electronic element 101 of the present invention is an electric double-layer transistor (EDLT) with an ion conductor 16 and a semiconductor layer 11 having a channel layer 15 as its core, and equipped with a gate electrode 14, a source electrode 12, and a drain electrode 13. The drain current of the EDLT is used as the reservoir. By performing reservoir calculations on hardware capable of high integration, it becomes possible to reduce computational resources. Here, the source electrode 12 and the drain electrode 13 are in electrical contact with the semiconductor layer 11. In particular, ohmic contact is preferred. Furthermore, as shown in Figure 1(b), the electric double-layer transistor 102 may have a structure in which a channel 15 is formed in a part of the surface layer of the semiconductor layer 11, and the source electrode 12 and drain electrode 13 are brought out on the same plane as the gate electrode 14, in order to facilitate integration.

[0012] The ion conductor 16 consists of a material that conducts ions and may be in solid, gel, or liquid form. If the ion conductor 16 is solid, it can be fabricated using semiconductor processes, making it easy to integrate, easy to handle, and free from leakage. If it is a gel, it has the advantage of reducing the risk of leakage and being relatively easy to handle. If it is a liquid, it has the advantage of easily adhering the ion conductor 16 to the semiconductor layer 11. The liquid may be organic or inorganic. For optimal reservoir properties, the thickness of the ion conductor 16 is preferably between 1 nm and 1 cm. Specific examples of ionic conductor 16 include Li-SiO2-ZrO2 (LZSO), PEO (polyethylene oxide) / LiClO4, and 1-ethyl-3-methylimidazolium-bis(trifluoromethanesulfonyl)imide (EMI-TFSI). Among these, LZSO is preferred because it is relatively stable in contact with reducing electrodes such as metallic Li.

[0013] There are no particular limitations on the conducting ions of ionic conductor 16, but Li + Ions, H + Ion, Na + Ions, Ag + Ions, Cu + Ions, Cl - Aeon, Br - Ion, I - Aeon, SO4 - Ions, FeCl4 - Ions, OH - Aeon, BF4 - Ions, PF annotation - AEON, CF3SO - Ion, (FSO2)2N - Ion, (CF3SO2)2N - Ion, C6H 11 N2 + Ion, C 10 H 20 F6N2O5S2 + One or more ions can be selected from the group consisting of ions. Among these, Li is widely used, has a large accumulation of know-how regarding its use, and also has the advantage of exhibiting relatively high ionic conductivity among solid materials. + Ions (Li ions) can be used preferentially.

[0014] Examples of semiconductor layers 11 include hydrogen-terminated diamond semiconductors, silicon, GaAs, Ga2O5, SiC, GaP, InP, ZnSe, CdS, GaN, SiGe, CuInSe2, graphene, monolayer MoS2, or ZnO. Among these, hydrogen-terminated diamond semiconductors are preferred because they have the characteristic of being resistant to degradation due to ion implantation (chemical stability). Here, single crystal diamonds are preferred. (111) diamonds and (100) diamonds can be preferred. Furthermore, silicon, which is widely used, high-quality, low-cost, and suitable for integration, can also be preferred as the semiconductor layer 11.

[0015] The channel layer 15 is formed on the surface of the semiconductor layer 11. Specifically, examples include the surface of a hydrogen-terminated diamond semiconductor (hydrogen-terminated portion) or a layer doped with impurities on the surface of a semiconductor such as silicon. Here, the channel layer 15 may consist of a single channel width or multiple channels of different widths. In the former case, where a single channel is present, it is necessary to convert the time-series data into multiple different input pulse voltage signals and input them repeatedly in order to obtain multiple reservoir states. In the case where multiple channels of different widths are present, it is possible to obtain multiple reservoir states simultaneously from a single input pulse voltage signal. Typical channel widths include those between 100 μm and 800 μm. In Figures 1(a) and (b), the channel layer 15 is in contact with the ion conductor 16, but direct contact is not necessarily required. A thin insulating film, such as SiO2, may be formed between the channel layer 15 and the ion conductor 16. The formation of a thin insulating film can suppress chemical reactions between the channel layer 15 and the ion conductor 16, thereby improving stability over time.

[0016] The gate electrode 14, source electrode 12, and drain electrode 13 can be made of any material that has low electrical resistance and does not react with the ion conductor 16, such as corrosion. Examples include metals selected from the group consisting of aluminum (Al), tungsten (W), gold (Au), titanium (Ti), platinum (Pt), palladium (Pd), copper (Cu), silver (Ag), magnesium (Mg), zinc (Zn), iron (Fe), cobalt (Co), nickel (Ni), and molybdenum (Mo), alloys containing one or more metals selected from the group, doped polysilicon, carbon, graphene, or graphite. Furthermore, the above materials can be used to form single-layer films or multilayer films combining the above materials. In Figures 1(a) and (b), the gate electrode 14, source electrode 12, and drain electrode 13 are in contact with the ion conductor 16, but direct contact is not necessarily required. A thin insulating film, such as SiO2, may be formed between them and the ion conductor 16. The formation of a thin insulating film can suppress chemical reactions between the gate electrode 14, source electrode 12, and drain electrode 13 and the ion conductor 16, thereby improving stability over time.

[0017] <Operation> Typical characteristics required for reservoir computing elements are (a) nonlinear properties, (b) short-term memory, and (c) high dimensionality.

[0018] In EDLT(101), high nonlinear characteristics are obtained from the complex response of the drain current flowing through the semiconductor channel, which is caused by both the change in electron carrier density of the semiconductor induced by the electric double layer at the ion conductor / semiconductor interface in response to the application of a pulsed voltage, and the charge-discharge behavior of the electric double layer. The complex drain current response used here arises from the series circuit formed by the electrical resistance (electron conduction) of the semiconductor, which depends on the charge state of the electric double layer, the capacitance of the electric double layer, and the ionic conduction resistance of the ion conductor, which does not depend on the charge state. Because it uses an electric double layer, the position and state of ions change over time, resulting in short-term memory properties. Furthermore, as shown in the examples, it also possesses high dimensionality. Furthermore, because it utilizes a phenomenon that occurs in an extremely thin region called the electric double layer, it is also suitable for high integration.

[0019] <Manufacturing method> EDLT(101) can be manufactured by the following process. First, a semiconductor layer 11 is prepared, and a source electrode 12 and a drain electrode 13 are formed on it. Methods for forming the source electrode 12 and drain electrode 13 include the lift-off method, and a method in which conductive material constituting the source electrode 12 and drain electrode 13 is deposited by sputtering or evaporation, and then processed into electrodes by lithography and etching. Subsequently, the ion conductor 16 is formed using methods such as CVD (Chemical Vapor Deposition), pulsed laser deposition, sputtering, or coating. Subsequently, the gate electrode 14 is formed. Methods for forming the gate electrode 14 include the lift-off method, and a method in which the conductive material constituting the gate electrode 14 is deposited by sputtering or vapor deposition, and then processed into an electrode by lithography and etching. Through the above process, the EDLT(101), an electronic component for reservoir calculations, is manufactured. The manufactured EDLT(101) possesses essential features for reservoir computing, including (a) nonlinear properties, (b) short-term memory, and (c) high dimensionality, and is compact, highly integrated, and can conserve computational resources.

[0020] (Embodiment 2) Embodiment 2 describes an electronic device for reservoir calculations that modulates semiconductor carriers by the ion transfer effect from an ion conductor to a semiconductor channel layer.

[0021] <Structure of electronic elements> The reservoir calculation electronic element 103 of Embodiment 2, as shown in Figure 8, is an ion transfer transistor (ITT) with an ion conductor 26 and a channel layer 21 as its core, and equipped with a gate electrode 14, a source electrode 12, and a drain electrode 13. The drain current of the ITT is used as the reservoir. By performing the reservoir calculation on hardware that can be highly integrated, it becomes possible to reduce the computational resources required. Here, the source electrode 12 and the drain electrode 13 are in electrical contact with the channel layer 21, which is a semiconductor layer. In particular, ohmic contact is preferred. There are no particular limitations on the ion conductor 26, but lithium ion conductive glass, lithium phosphate, Nafion, mesoporous silica, Rb, Ag4I5, Rb4Cu 16 I7Cl 13 AgI, CuI, Li3N, Li-SiO2-ZrO2(LZSO), PEO(polyethylene oxide) / LiClO4, 1-ethyl-3-methylimidazolium-bis(trifluoromethanesulfonyl)imide (EMI-TFSI) can be used as the channel layer 21, and WO3, TiO2, FeO4, LiCIO2, LiFePO4, LiNiO2, Li2MnO3, Na2Fe2(SO4)3, Na7V(4PO4), P2O7, K 2 / 3 Ni 2 / 3 Te 1 / 3 O2, MoS2, ZnO, graphene, graphite, graphite oxide, Ag2S, Cu2S can be used. There are no particular limitations on the ions that can be moved, but Li + Ions, H + Ion, Na + Aeon, K + Ions, Ag + Ions, Cu + Ions, Cl - Aeon, Br - Ion, I - Aeon, SO4 - Ions, FeCl4 - Ions, H - Ions, OH - Aeon, BF4 - Ions, PF annotation -AEON, CF3SO - Ion, (FSO2)2N - Ion, (CF3SO2)2N - Ion, C6H 11 N2 + Ion, C 10 H 20 F6N2O5S2 + You can list one or more ions selected from the group consisting of ions. As shown in Example 2, ITT(103) possesses essential functions for reservoir computation, including (a) nonlinear properties, (b) short-term memory, and (c) high dimensionality, and is compact, highly integrable, and can conserve computational resources.

[0022] In the EDLT(101) of Embodiment 1, the case in which the drain current is used as the reservoir state was described, but the target of use as the reservoir state is not limited to the drain current. As shown in Embodiment 2 below, the reservoir element of the present invention can also use the gate current as the target of use as the reservoir state.

[0023] Furthermore, as shown in Figure 11, the information processing device 1001 may be configured such that the input signal from the input element 301 that processes input information is processed by the reservoir element 302 of the present invention (a reservoir computing electronic element described in any one of the above configurations 1 to 11), the output signal is transmitted to the output element 303, and the output element 303 processes and outputs the result. The information processing device 1001 can be configured to be an information processing device that requires few computing resources, consumes little power, and is also compact in size. [Examples]

[0024] (Example 1) In Example 1, an electric double layer transistor (EDLT) that modulates semiconductor carriers through the electric double layer (EDL) effect at the ion conductor-semiconductor interface was fabricated, physically implemented as a reservoir computing electronic element, and its performance was evaluated through tasks such as image recognition and time-series data analysis.

[0025] <Elements and their structure> The main structural components of EDLT(101) are shown in Figures 1(a) and 2. The ion conductor 16 was a Li-SiO2-ZrO2(LZSO) thin film, which is a Li ion conductor deposited by pulsed laser deposition. Its thickness was approximately 700 nm. As the semiconductor layer (substrate) 11, a single-crystal diamond with a thickness of 300 μm and a crystal plane of (100) was used. The channel layer 15 used hydrogen-terminated diamond, which is chemically inert and exhibits an ideal EDL effect. Here, the hydrogen-terminated diamond was fabricated by chemical vapor deposition. Specifically, it was fabricated by introducing 0.5 sccm of methane and 1000 sccm of hydrogen, respectively, using microwave plasma chemical vapor deposition with an RF power of 950 W and a substrate temperature of 940 °C. The channel length is 100 μm and the channel width is 800 μm. Both the source 12 and drain 13 electrodes are made of palladium and platinum, which were continuously deposited by electron beam evaporation, with thicknesses of 10 nm and 35 nm, respectively. The gate electrode 14 consists of lithium cobalt oxide deposited by pulsed laser deposition and platinum deposited by electron beam deposition, with thicknesses of 100 nm and 50 nm, respectively.

[0026] <Method of characterization> The input data is applied to the gate electrode 14 as an input voltage signal, and the drain current response I at that time is measured. D This was used as the reservoir output (Figure 2). The operating characteristics of the EDLT are determined by the ions Li in the ion conductor. + This is determined by the behavior of electrons in hydrogen-terminated diamond, but as shown below, good input-output nonlinearity and short-term memory were confirmed. The measurement device used was a 4200A-SCS (manufactured by Keithley).

[0027] <Characteristic Evaluation 1> Characteristic evaluation 1 is an example of handwritten character recognition performed using EDLT101 as a reservoir. There, the 28x28 pixel handwritten digits are converted into binary information of "1" and "0" depending on whether input is present or not for each pixel, then converted into time-series data every 4 pixels and input to EDLT(101), and the output current I from the drain electrode 13 is generated. D We measured it.

[0028] In detail, the measurements were performed using the procedure shown below. The time-series data consists of 16 possibilities, from "0000" to "1111," which were converted into pulse voltage signals. The pulse interval of the pulse voltage signal is 12.5 ms, and the pulse width and pulse amplitude are 10 ms and 0.8 V, respectively. Then, while applying a drain voltage of -0.5 V to the drain electrode 13, the pulse voltage signal was applied to the gate electrode 14, and the drain current response I at that time was measured. D We measured it.

[0029] For example, when time-series data corresponding to the "1011" pixels is input, the drain current I D The response is shown in Figure 3. From this result, the drain current I, which is the output relative to the input, can be seen. D It can be seen that it changes non-linearly and has short-term memory characteristics that reflect past inputs.

[0030] Next, the reservoir state was defined as the drain current after 4 data inputs, normalized so that the drain current value after inputting "1111" becomes 1. Figure 4 shows the reservoir state when 16 different pixel information values ​​from "0000" to "1111" are input. Since the reservoir state takes on a different value for each of the 16 inputs, it can be seen that the system classifies the 16 types of inputs well.

[0031] Next, the reservoir state was obtained when the sign of the input pulse voltage signal was inverted. This state was designated as channel 1 if it was not used for image recognition, and channel 2 if it was used. In image recognition, 784-pixel image data was converted into data consisting of either 196 data points (1 channel) or 392 data points (2 channels) of reservoir states, and this was input into the network of the reading unit for training and testing. The network structure of the readout section is a 196×10 or 392×10 network, consisting of an input layer with 196 or 392 nodes corresponding to the number of input data points and an output layer with 10 nodes corresponding to the digits 0 through 9. Here, the sigmoid function was used as the activation function for the output layer.

[0032] The network was trained using the least squares method with a learning rate of 0.1, and the image recognition accuracy was calculated by inputting 10,000 test data points into the trained network and analyzing the correct / incorrect results. Figure 5 shows the results of evaluating the accuracy rate while varying the number of training images from 100 to 60,000. For reference, the image recognition accuracy rate when using a 3-layer neural network commonly used in image recognition is also shown. The neural network structure is a 784 × 100 × 10 fully connected network with 784 neurons in the input layer, 100 neurons in the hidden layer, and 10 neurons in the output layer. The learning rate from the input layer to the hidden layer is 0.4, and the learning rate from the hidden layer to the output layer is 0.2. Furthermore, the sigmoid function was used as the activation function, and learning was performed using backpropagation.

[0033] Image recognition accuracy using reservoir computation with EDLT(101) showed accuracy comparable to a 3-layer neural network in regions where the number of training images was 2000 or less, which is smaller than the number of test images, i.e., regions with higher recognition difficulty. While a 3-layer neural network requires 784 × 100 × 10 = 784,000 inter-node links for training, the reservoir computation requires only 196 × 1 = 1960 inter-node links for 1 channel and 392 × 10 = 3920 links for 2 channels, which is 1 / 400th and 1 / 200th of the neural network, respectively. Despite being a much smaller network than a 3-layer neural network, the fact that it achieved a comparable accuracy demonstrates the high capability of this device as a reservoir.

[0034] <Characteristic Evaluation 2> Characterization 2 presents an analysis task involving quadratic nonlinear dynamical equations. Let the input be a time series data u(κ) which takes random values ​​from 0 to 0.5. y(κ)=0.4y(κ−1)+0.4y(κ−1)y(κ−2) +0.6u 3 (κ) + 0.1 ···Equation (1) We studied and tested the u-to-y transformation of the quadratic nonlinear dynamical equations represented by . Note that κ is discrete time.

[0035] An example of the input waveform u(κ) and y(κ) represented by equation (1) is shown in Figure 6. The input data u(κ) was converted into pulse voltage signals (pulse width 10 ms, pulse amplitude 0.8 V) with seven different pulse intervals (1 ms, 2 ms, 5 ms, 10 ms, 20 ms, 50 ms, 100 ms) and input to EDLT(101), and the predicted waveform was regenerated by a linear combination of the 7 channels of output data. Therefore, the network size of the readout section is 7 × 1. Then, the reading unit was trained using linear regression based on the training data obtained from equation (1). Next, random waves u(κ) different from the training data were input to EDLT(101) as test data, and the reservoir state was obtained. The waveform obtained by inputting these reservoir states into the trained network was used as the predicted waveform, and the error was calculated by comparing it with the target waveform obtained from equation (1) to evaluate the accuracy.

[0036] Figure 7 shows the predicted and target waveforms obtained from the test data. The normalized mean squared error (NMSE) at this time is 2.77 × 10⁻⁶. -3 The figures were sufficiently low, showing good agreement. Equation (1) is a nonlinear equation with a first-order lag in the first term on the right-hand side and a second-order lag in the second term on the right-hand side. The good agreement indicates that the reservoir using EDLT(101) has good nonlinear and short-term memory characteristics. It also shows that the input waveform is made high-dimensional enough to perform this task through pulse interval control.

[0037] (Example 2) In Example 2, an ion-transfer transistor (ITT) (103) that modulates semiconductor carriers by the ion transfer effect from an ion conductor to a semiconductor channel layer was fabricated, physically implemented as a reservoir computing electronic element, and its performance was evaluated through a time-series data task, similar to Example 1.

[0038] <Elements and their structure> The main structure is shown in Figure 8. Ion conductor 26 was a lithium-conducting glass substrate (manufactured by Ohara Corporation), which is a Li-ion conductor. Its thickness was 150 μm. The channel layer 21 used a WO3 thin film that induces reversible ion transfer. Here, the WO3 was fabricated by AC sputtering. The channel length is 100 μm and the channel width is 800 μm. Both the source 12 and drain 13 electrodes are made of platinum deposited by AC sputtering, with a thickness of 60 nm each. The gate electrode 14 consists of lithium cobalt oxide deposited by pulsed laser deposition and platinum deposited by electron beam deposition, with thicknesses of 100 nm and 50 nm, respectively.

[0039] <Method of characterization> The input data is applied to the gate electrode 14 as an input voltage signal, and the drain current response I at that time is measured. D , and gate current response I GThis was used as the reservoir output (Figure 8). The operating characteristics of ITT(103) are mainly determined by the behavior of ions and electrons in WO3 and ions in the ion conductor, but as shown below, good input-output nonlinearity and short-term memory were confirmed. A 4200A-SCS (Keithley) was used as the measurement device.

[0040] <Characteristic Evaluation> The characterization section demonstrates the analysis task of quadratic nonlinear dynamical equations. Similar to Example 1, the input is a time series data u(κ) that takes random values ​​from 0 to 0.5. y(κ)=0.4y(κ−1)+0.4y(κ−1)y(κ−2) +0.6u 3 (κ) + 0.1 ···Equation (1) We studied and tested the u-to-y transformation of the quadratic nonlinear dynamical equations represented by . Note that κ is discrete time.

[0041] The input data u(κ) was converted into a pulse voltage signal with a pulse interval of 10s (pulse width 10s, pulse amplitude 0.5V) and input to ITT(103). The predicted waveform was regenerated using a linear combination of the output data of 20 drain current channels, and a total of 40 channels, which consist of the output data of 20 gate current channels in addition to the 20 drain current channels. Therefore, the network size of the readout section is 20×1 or 40×1. Then, the reading unit was trained using linear regression based on the training data obtained from equation (1). Next, a random wave u(κ) different from the training data was input to ITT(103) as test data, and the reservoir state was obtained. The waveform obtained by inputting these reservoir states into the trained network was used as the predicted waveform, and the error was calculated by comparing it with the target waveform obtained from equation (1) to evaluate the accuracy.

[0042] Figure 9 shows the predicted and target waveforms obtained from test data using only the drain current. The normalized mean squared error (NMSE) in this case is 7.78 × 10⁻⁶. -4The result was sufficiently low, showing good agreement. Figure 10 shows the predicted and target waveforms obtained from test data using both drain current and gate current. The normalized mean squared error (NMSE) in this case was 5.79 × 10⁻⁶. -4 This resulted in improved prediction accuracy compared to the case using only the drain current. Equation (1) is a nonlinear equation with a first-order lag in the first term on the right-hand side and a second-order lag in the second term on the right-hand side. The good agreement indicates that the reservoir using ITT(103) has good nonlinear and short-term memory characteristics. It also shows that the input waveform is made high-dimensional enough by pulse interval control to perform this task. [Industrial applicability]

[0043] The present invention provides an electronic element for reservoir calculations that exhibits high nonlinearity in the reservoir section, thus reducing computational resources, and is compact and suitable for integration. Therefore, we believe that the present invention will greatly contribute to industrial development. [Explanation of Symbols]

[0044] 11: Semiconductor layer (hydrogen-terminated diamond) 12: Source electrode (source) 13: Drain electrode (drain) 14: Binding gate 15: Channel Layer 16: Ionic conductor (LSZO) 21: Channel layer (WO3) 26: Ion conductor (lithium-ion conductive glass substrate) 101: Electronic components (reservoir elements, electric double-layer transistors, EDLTs) 102: Electronic components (reservoir elements, electric double-layer transistors, EDLTs) 103: Electronic devices (reservoir elements, ion transfer transistors, ITTs) 301: Input element 302: Reservoir element 303: Output element 1001: Information Processing Device

Claims

1. a semiconductor layer, an ion conductor, a gate electrode, a source electrode, and a drain electrode; the source electrode and the drain electrode are in electrical contact with the semiconducting layer; The gate electrode is disposed in a channel layer formed between an electrical contact surface of the source electrode and an electrical contact surface of the drain electrode on the semiconductor layer, via the ion conductor.

2. The semiconductor layer may be a hydrogen-terminated diamond semiconductor, silicon, GaAs, Ga ２ O ５ , SiC, GaP, InP, ZnSe, CdS, GaN, SiGe, CuInSe ２ , graphene, single-layer MoS ２ 2. The reservoir computational electronic device of claim 1, comprising one selected from the group consisting of: ZnO, and ZnO.

3. The reservoir computational electronic device of claim 2 , wherein the semiconductor layer is made of a hydrogen-terminated diamond semiconductor.

4. The conductor ion of the ionic conductor is Li ＋ Ion, H ＋ Ion, Na ＋ Ion, Ag ＋ Ion, Cu ＋ Ion, Cl － Ion, Br － Ion, I － Ion, SO ４ － ion, FeCl ４ － ion, OH － Ion, BF ４ － Ion, PF ６ － Ion, CF ３ SO － ion, (FSO ２ ) ２ N － ion, (CF ３ SO ２ ) ２ N － Ion, C ６ H １１ N ２ ＋ Ion, C １０ H ２０ F ６ N ２ O ５ S ２ ＋ 4. The reservoir computational electronic element according to claim 1, wherein the reservoir computational electronic element is one or more selected from the group consisting of ions.

5. The conductor ion of the ionic conductor is Li ＋ 5. The reservoir computational electronic element of claim 4, which is an ion.

6. 2. The reservoir computational electronic device of claim 1, wherein the ionic conductor is solid-state.

7. 2. The reservoir computational electronic device of claim 1, wherein the ionic conductor is in the form of a gel.

8. 2. The reservoir computational electronic device of claim 1, wherein the ionic conductor is in a liquid state.

9. The ion conductor is Li—SiO ２ -ZrO ２ 2. The reservoir computational electronic device of claim 1, comprising: (LZSO).

10. 2. The reservoir computational electronic device of claim 1, wherein the gate electrode, the source electrode, and the drain electrode are made of one or more selected from the group consisting of a metal selected from the group consisting of aluminum (Al), tungsten (W), gold (Au), titanium (Ti), platinum (Pt), palladium (Pd), copper (Cu), silver (Ag), magnesium (Mg), zinc (Zn), iron (Fe), cobalt (Co), nickel (Ni), and molybdenum (Mo), an alloy containing one or more metals selected from the group, doped polysilicon, carbon, graphene, and graphite.

11. The reservoir computational electronic device according to claim 1 , wherein the channel layer has a plurality of channels each having a different channel width arranged therein.

Images